Most UK accountancy practices that stall on AI are not held back by a lack of enthusiasm. They are held back by a lack of structure. Without a clear sequence of steps, even well-motivated teams end up in the same place: tools installed but barely used, no agreed policy, and no one quite sure who is responsible for what.
This article is part of Runbook's complete guide to AI implementation for UK accountancy practices. It sets out the checklist every practice needs before, during, and after introducing AI tools: from assessing your current position to getting your team using AI consistently and safely. If you want to find out where your practice stands right now, the free AI Readiness Scorecard gives you a personalised picture in under five minutes.
Before you start: the readiness assessment
Jumping straight to tool selection is the most common implementation mistake. Practices that start by asking "which AI tool should we use?" before they have answered a more fundamental question, "what are we actually trying to solve?", tend to end up with tools that do not get used.
The readiness assessment is a short internal audit that takes no more than an hour and covers four areas.
1. Identify your highest-volume, lowest-risk tasks
List the tasks in your practice that are high volume, involve writing or document handling, and where the output is always reviewed by a qualified person before it reaches a client or a regulatory body. These are your AI candidates. Common examples include drafting routine client correspondence, summarising meeting notes, producing first drafts of covering letters, and researching HMRC guidance.
Tasks involving direct client advice, statutory filings, or complex calculations are not your starting point. They may become AI-assisted in time, but they are not where you begin.
2. Assess your current data handling practices
Before any tool is selected, you need a clear picture of what types of data flow through your practice and how it is currently stored and processed. This matters because different AI tools have very different data handling policies, and some of the most widely used consumer tools are not suitable for processing identifiable client information. Knowing your data landscape before choosing a tool prevents expensive mistakes later.
3. Gauge staff baseline and appetite
Find out who in your team is already using AI tools informally, and who is sceptical or uncertain. Both groups matter. Informal users represent your early adopters and likely trainers; sceptics often have legitimate concerns about quality and risk that your implementation process needs to address. A short survey or team conversation at this stage is time well spent.
4. Set a realistic scope for phase one
Phase one should cover one or two tasks, one approved tool, and a defined group of users (often three to five people). Anything larger becomes difficult to manage, harder to measure, and more likely to stall. Write down what success looks like after 90 days. That is your target.
Practical example: A 12-person practice might start its AI implementation by identifying that fee earners were spending an average of 25 minutes per client meeting writing up notes and action points. Phase one would be limited to that single task, using Microsoft Copilot with a Teams licence, with three senior members of staff. After six weeks, they would find that note-writing time had reduced to under five minutes per meeting across the group, and the process was documented well enough to roll out to the rest of the team.
Selecting the right tools for your practice
Tool selection is step three in the process, not step one. Once you know what you are trying to achieve and what your data constraints are, the shortlist of suitable tools becomes much shorter.
The main options for UK practices in 2026
For most UK accountancy practices, the realistic shortlist comes down to three options for general-purpose AI assistance.
- Microsoft Copilot with a Microsoft 365 Business licence. The most natural choice for practices already on the Microsoft 365 stack. Copilot integrates with Outlook, Word, Teams, and Excel, and the business licence includes a data processing agreement suitable for professional use. The limitation is that it works best within Microsoft applications; it is less flexible for tasks outside that environment.
- ChatGPT Team. The paid business tier of ChatGPT, which includes a data processing agreement and does not use conversation data to train the model. More flexible than Copilot for open-ended writing and research tasks. Requires users to work in the ChatGPT interface rather than inside existing tools, which can affect adoption.
- Claude for Work (Anthropic). An alternative to ChatGPT for writing-heavy tasks, with strong performance on long-form drafting, document summarisation, and instruction-following. Business tiers include appropriate data handling agreements.
Purpose-built AI features inside your existing accountancy software (Xero, QuickBooks, Sage) are a separate category and worth assessing independently. These are designed specifically for accounting data and operate within the software provider's own compliance framework, making them a lower-risk starting point for data-heavy work.
What to check before approving any tool
- Does the provider offer a UK GDPR-compliant data processing agreement on the tier you are purchasing?
- Is input data used to train or improve the model? (It should not be, for any tool handling client information.)
- Where is data stored and processed? (EU or UK data residency is preferable for UK practices.)
- Does the tool integrate with your existing workflow, or does it require staff to change their working environment significantly?
The Runbook AI Implementation Checklist covers every stage in detail, including a tool evaluation scorecard, a data protection checklist, a staff rollout template, and a 90-day implementation plan. Designed specifically for UK accountancy practices.
Data protection and GDPR: the non-negotiable stage
This is the stage most practices underestimate, and the one where getting it wrong has the most serious consequences. UK GDPR applies to any processing of personal data, which includes client names, contact details, financial information, and any data that could identify an individual. If you are inputting any of this into an AI tool, the rules apply.
Write a data processing policy for AI tools
Before any tool is used with client information, your practice needs a written policy that covers the following.
- Which tools are approved and on which tier (free tiers are almost never suitable for client data).
- What categories of data can be inputted into each tool, and what cannot.
- Who has approved access to each tool.
- What the review requirement is for any AI output that will be sent to a client or used in a filing.
This policy does not need to be lengthy. A single page covering these four points is sufficient to start. It should be reviewed when you add new tools or change tiers on existing ones.
Confirm your data processing agreements are in place
A data processing agreement (DPA) is a contract between your practice and the AI tool provider that sets out how your data is handled. Under UK GDPR, you are required to have a DPA in place with any third party that processes personal data on your behalf. Most paid business tiers of the major AI tools include a DPA, but you should check the specific tier you are on and keep a record that the agreement is in place.
Important: The free tiers of ChatGPT, Claude, and most other consumer AI tools do not include data processing agreements and typically use input data to improve the model. They are not suitable for processing identifiable client information. If any member of your team is currently using a free AI tool for client work, this needs to be addressed before your implementation proceeds.
Consider your client notification obligations
If AI tools are processing client data as part of your service delivery, you may need to update your privacy notice to reflect this. Your data protection adviser or legal team is the right resource for advice specific to your circumstances. Runbook does not provide GDPR or legal advice, but we recommend raising this question before your implementation is complete rather than after.
Staff rollout and training
The difference between a practice where AI gets used consistently and one where it gets used occasionally by a few enthusiasts almost always comes down to how the rollout was handled. Tools introduced without training, context, or a clear reason for adoption tend not to get adopted.
Phase the rollout
Start with a small group of willing participants, ideally three to five people who handle the tasks you have identified as phase one priorities. Give this group four to six weeks to develop a working approach before expanding. The goal at this stage is not widespread adoption; it is a tested, documented process that can be handed to the wider team with confidence.
Provide structured prompts, not just access
Giving staff access to an AI tool and telling them to get on with it is not training. The most common reason AI tools fail to deliver time savings is that staff do not know how to write effective prompts. Providing a short library of tested prompts for the specific tasks you want AI to help with produces faster adoption and more consistent results than open-ended experimentation.
The 20 ChatGPT prompts for UK accountants article includes ready-to-use prompts for common accountancy tasks, including client correspondence, meeting notes, and HMRC-related writing. These are a practical starting point for building your own prompt library.
Set clear expectations about review
Every member of staff using AI tools needs to understand that AI output is a draft, not a finished product. This applies to client emails, internal documents, meeting summaries, and everything else. Building a review step into every AI-assisted workflow is not optional. It is the safeguard that makes AI use safe and reliable.
Document your agreed workflows
Once your phase one group has developed a working approach for each task, write it down. A workflow does not need to be elaborate: a short description of the task, the prompt structure that works best, and the review process before the output is used is sufficient. Documented workflows make it possible to train new staff, maintain consistency, and build on what is working rather than starting from scratch each time.
Client communication
Whether and how you communicate your AI use to clients is a question practices handle differently. There is no single right answer, but there are considerations worth working through before your implementation is complete.
When to tell clients
If AI is being used to draft correspondence that is sent to clients, most practices take the view that this falls within the scope of normal professional tools, in the same way that using document assembly software or practice management systems does not require specific client notification. If AI is being used in a way that directly affects the substance of advice or output (for example, using AI to generate the first draft of a tax planning recommendation), a more explicit conversation with the client may be appropriate.
Your professional body's guidance is the primary reference point here. The ICAEW and ACCA have both published guidance on AI in professional practice, and this should inform your approach.
What to do when clients ask
Prepare a brief, honest answer to the question "do you use AI?" that you and your team can give consistently. Something straightforward works best: you use AI tools to assist with drafting and administrative tasks; all output is reviewed by a qualified member of the team before it is used or sent; you do not input identifiable client data into any tool without a data processing agreement in place. Most clients who ask will be satisfied with a clear, honest answer. Evasion or vagueness tends to produce more concern, not less.
Review, measure, and iterate
An AI implementation that is not reviewed is an AI implementation that drifts. After 90 days, your phase one process should be evaluated against the targets you set at the outset.
What to measure
- Time saved per task. Compare the time the task took before AI assistance with the time it takes now, including the review step. If the saving is minimal, the prompt or workflow needs adjusting, not the tool.
- Error rate and quality. Are AI outputs requiring heavy editing? If so, the prompt structure probably needs improvement. Are any errors reaching clients? If so, the review process needs strengthening.
- Adoption rate. Are the people who were supposed to use the tool actually using it? Low adoption after a structured rollout usually points to a training gap or a workflow that has not been sufficiently documented.
When to expand
The signal to move to phase two is a phase one process that is working reliably and is documented clearly enough to hand to new users without significant hand-holding. Once that threshold is met, you can add tasks, add users, or both. Do not expand until phase one meets that standard. Expanding a process that is not yet reliable simply multiplies the inconsistency.
For a structured framework covering every stage of this process in detail, including templates, a tool evaluation scorecard, and a 90-day rollout plan, the AI Implementation Checklist for UK Accountancy Practices covers the full process from readiness assessment to ongoing review.
Frequently asked questions
How long does AI implementation take for a small accountancy practice?
A focused implementation covering one or two tasks can produce visible results within four weeks. A fuller rollout covering tool selection, data policy, staff training, and client communication typically takes 60 to 90 days when done properly. Practices that try to do everything at once tend to see slower adoption than those who follow a phased approach.
Do we need a written AI policy before we start using AI tools?
Yes, if you have any members of staff using AI tools, or if you intend to process any client-related information. A written policy does not need to be lengthy, but it should cover which tools are approved, what data can and cannot be inputted, and who is responsible for reviewing AI output. Without a policy, you have no way to ensure consistency or manage risk across the team.
Which AI tools should a UK accountancy practice start with?
For most practices, a sensible starting point is either ChatGPT Team or Microsoft Copilot with a Microsoft 365 Business licence, both of which include data processing agreements suitable for professional use. The right choice depends on your existing software environment and the tasks you want to tackle first. Tool selection is step three in the implementation process, not step one.
What data can we put into AI tools under UK GDPR?
Identifiable client data should not be entered into any AI tool unless you have confirmed that the provider offers a UK GDPR-compliant data processing agreement, and that data is not used to train the model. The free tiers of most consumer AI tools do not meet this requirement. Paid business tiers from the major providers typically do, but you should verify this with your data protection adviser before proceeding.
What is the biggest mistake practices make when implementing AI?
Trying to do too much at once. Evaluating multiple tools simultaneously, attempting to automate complex workflows before simpler ones are working, and involving the whole team before anyone has developed real proficiency all tend to produce confusion rather than adoption. The most reliable path is to start with one task, one tool, and a small group, and expand from there once you have a working process.